Why Your AIOps Agents Are Only As Good As Your Infrastructure Data (+ OpsMill Infrahub Review)
AIOps deployments fail because of bad data, not bad algorithms. How graph-based platforms like Infrahub fix the infrastructure data problem feeding your AI agents.
TLDR: Most AIOps failures trace back to a data problem, not an algorithm problem. When AI agents act on stale CMDBs, disconnected spreadsheets, and ad-hoc scripts, errors cascade fast. Graph-based infrastructure data platforms like OpsMill’s Infrahub address this by modeling real infrastructure relationships rather than flattening everything into asset tables. Infrahub’s open-core model (free Community edition, licensed Enterprise) makes it accessible to evaluate, but the Enterprise edition’s pricing is opaque and requires a sales conversation. If your org is scaling AI-driven ops beyond a single domain, fix the data layer first or accept that your agents will keep flying blind.
Why This Matters Now
OpsMill announced a $14M Series A today (May 7, 2026), led by IRIS with BGV, Serena, and Partech participating. The round is notable less for the dollar amount than for the thesis behind it: the company’s bet is that the biggest bottleneck in enterprise AIOps adoption is not the AI layer but the infrastructure data feeding it.
The numbers support this framing. Gartner forecasts that 30% of enterprises will automate more than half their network activities by 2026, up from under 10% in mid-2023. ITIC’s 2024 downtime report pegs the average cost of an hour of unplanned outage at $300,000. Yet the data describing most enterprises’ infrastructure — what exists, how it connects, how it should be configured — remains scattered across tools that were never designed for agent consumption. ABI Research estimates the broader AI-driven IT operations market at $220 billion, and demand for enterprise network automation has tripled since 2023. The infrastructure data layer is the foundation nobody wants to talk about.
The Infrastructure Data Problem at a Glance
| Dimension | Traditional CMDB | Spreadsheets / Scripts | Graph-Based Platform (e.g., Infrahub) |
|---|---|---|---|
| Data model | Flat tables, predefined fields | Unstructured, per-team conventions | Flexible schema, relationship-first |
| Relationship mapping | Limited; foreign keys at best | None | Native graph connections |
| Version control | Snapshot-based or none | Manual file versioning | Git-integrated branching, diff, merge |
| Change validation | Manual review or post-deploy audit | None | Built-in CI pipelines pre-merge |
| AI/agent readiness | Requires ETL middleware | Requires custom parsing | GraphQL/REST APIs, MCP server for LLM agents |
| Typical staleness | Days to weeks behind reality | Instantly stale once someone forgets to update | Near-real-time via sync connectors |
| Scale ceiling | Enterprise-grade but rigid | Breaks past a few hundred assets | Graph DB scales to hundreds of thousands of elements |
| Cost model | Per-node or per-user licensing (often $50-200+/node/year) | Free but unmaintainable | Open-source Community; Enterprise pricing undisclosed |
The Hidden Reason AIOps Deployments Fail
The AIOps conversation in 2026 overwhelmingly focuses on the AI layer: which platform runs better anomaly detection, whose incident correlation is faster, which LLM agent can auto-remediate a P1. Almost nobody talks about what those agents are actually reading.
The typical enterprise infrastructure data stack looks like this: a CMDB that was configured once during a migration project and has drifted since, a collection of team-specific spreadsheets that contradict each other, and a set of scripts that scrape configuration from live devices but store it in formats only the author understands. When an AIOps agent queries this mess for the current state of a firewall rule or the dependency chain behind a microservice, it gets incomplete or conflicting answers. The agent doesn’t know the data is bad. It acts on it anyway.
In financial services, a misconfigured firewall rule surfaced by stale CMDB data can trigger a regulatory penalty. In manufacturing, an agent deploying a network change based on an outdated topology map can shut down a production line. The failure mode is not “AI made a bad prediction.” The failure mode is “AI confidently executed a correct action on incorrect data.”
Earned insight: In my experience, the single most common root cause of failed AIOps pilot projects is not model accuracy or integration complexity — it is that the infrastructure data the AI consumed was between two weeks and six months out of date. Teams spend months tuning correlation rules when they should have spent weeks auditing their data layer.
What a Graph-Based Approach Changes
Traditional CMDBs store infrastructure as rows in tables: one row per server, one row per switch, with foreign keys linking them when someone remembers to set them up. This flat model works for asset accounting but collapses when you need to answer questions like “what services depend on this switch, and what happens if it goes down at 2 AM?”
A graph database stores infrastructure elements as nodes and their relationships as edges. A server is connected to a VLAN, which is connected to a firewall policy, which is connected to a compliance requirement. The relationships are first-class data, not afterthoughts. When an AI agent needs to reason about blast radius or dependency chains, the answers are already in the topology — no multi-table joins, no ETL pipelines, no guessing.
This matters specifically for AIOps because modern incident response is relationship-dependent. An alert about high latency on a pod means nothing without the context of which service runs on that pod, which database it connects to, which network path it traverses, and whether any of those components changed recently. Graph-native platforms make that context queryable by default.
OpsMill Infrahub: What It Does
Infrahub is an open-source infrastructure data management platform built on Neo4j (graph database), Git, and Python. Founded in 2023 by Damien Garros (ex-Juniper, ex-Roblox, ex-Network to Code) and Karen Gallantry, the platform is designed as a single source of truth for infrastructure teams and the AI agents working alongside them.
What Works
Flexible, schema-first data modeling. Unlike CMDBs that force you into predefined fields, Infrahub lets you define your own data model — devices, services, topologies, business context — enforced with validation. This means the model matches your actual infrastructure, not a vendor’s assumption of what infrastructure looks like.
Git-integrated version control for data. Every change to schema, data, or generated configurations goes through a branch-diff-review-merge workflow identical to code review. This is a genuine differentiator: CMDBs track snapshots at best, and most track nothing. Infrahub gives you an immutable, queryable history of who changed what, when, and why.
Built-in CI validation. Proposed changes run through CI pipelines before merge. If a change would break a dependency or violate a constraint, it fails before reaching production. This is the kind of safety net that turns infrastructure data from “probably right” into “validated right.”
AI-ready integrations. Infrahub ships with a Model Context Protocol (MCP) server, making its validated data directly consumable by Claude, Cursor, and other LLM-based agents. It also includes GraphQL and REST APIs, plus sync connectors for NetBox, Nautobot, and IP Fabric.
Real customer traction. TikTok runs Infrahub in production. Eurofiber, a European cloud services provider, cut deployment times from five days to fifteen minutes after implementing it. The open-source community includes hyperscalers, global retailers, fintechs, and insurers.
Where It Struggles
Pricing opacity. The Enterprise edition pricing is entirely hidden behind a “contact sales” wall. The pricing page mentions “processing capabilities” as the pricing axis but discloses no numbers. For a startup competing against well-documented alternatives, this is a friction point. Early-stage enterprise buyers want to model TCO before scheduling a demo.
Network/infrastructure focus, not full-stack. Infrahub is purpose-built for network and data center infrastructure. If you need a unified data platform that also covers application-layer services, SaaS dependencies, or business process mapping, you will need to supplement Infrahub with other tools or rely on its sync connectors.
Early-stage ecosystem. Founded in 2023, the company is still building out its partner ecosystem, integrations, and documentation. The Discord community is active but small compared to alternatives like NetBox (which has years of community momentum). Enterprise features like SSO validation, high-availability Neo4j, and hardened deployments are Enterprise-only.
Competitive positioning requires nuance. OpsMill positions against CMDBs and GitOps, but the more direct competition is NetBox Labs and Nautobot — tools that many network teams already know and use. Migrating from a working NetBox instance to Infrahub requires a clear ROI case, and OpsMill’s marketing does not yet make that case sharply.
Warning: Infrahub’s Community edition uses Neo4j Community, which lacks high-availability clustering. For production deployments where infrastructure data availability is itself critical, you will need the Enterprise edition — and that means entering a sales cycle with undisclosed pricing. Factor this into your evaluation timeline.
Infrahub Strengths:
- Graph-native data model maps real infrastructure relationships, not flat tables
- Git-integrated version control for all data and schema changes
- Built-in CI validation prevents bad data from reaching production
- MCP server provides direct LLM/agent integration out of the box
- Open-source Community edition under Apache 2.0 — no cost to evaluate
- Real production traction at TikTok and Eurofiber
Infrahub Weaknesses:
- Enterprise pricing completely undisclosed; no public TCO guidance
- Focused on network/data center infrastructure — not a full-stack CMDB replacement
- Smaller community and ecosystem than established alternatives like NetBox
- High-availability deployment requires Enterprise edition
- Early-stage company (Series A) — long-term viability not yet proven at scale
Who Else Plays in This Space
Infrahub does not compete in a vacuum. The infrastructure data management landscape includes several alternatives worth evaluating depending on your starting point:
ServiceNow CMDB remains the enterprise default, deeply integrated with ITSM workflows and compliance reporting. It is the most mature option but also the most rigid — adapting it for AI agent consumption typically requires middleware and significant configuration. Pricing runs $50-100+/user/month depending on the module stack.
NetBox Labs (NetBox) is the most widely adopted open-source network source of truth, with a large community and extensive plugin ecosystem. It uses a relational database, not a graph, which limits relationship querying. NetBox Cloud offers a hosted version for teams that don’t want to self-manage.
Nautobot forked from NetBox and adds a job/automation framework. It is a strong choice for teams already embedded in the Ansible/Python automation ecosystem, though it shares NetBox’s relational database limitations.
Itential focuses on network automation orchestration — more of an execution layer than a data layer. It can consume data from CMDBs or source-of-truth tools but does not replace them.
Port.io and Backstage serve as internal developer portals and service catalogs. They map application-layer context well but generally lack the deep infrastructure topology modeling that AIOps agents need.
Earned insight: The choice between a graph-based platform and a relational one often comes down to query complexity. If your agents only need “what IP is assigned to this server,” a relational source of truth is fine. If they need “show me every service impacted if this switch fails, including the compliance implications,” you need a graph. Most teams underestimate how quickly their query needs escalate once agents start running autonomously.
Pricing Reality
This is where transparency breaks down across the category:
| Platform | Community / Free Tier | Enterprise Cost | Pricing Model | Hidden Costs |
|---|---|---|---|---|
| Infrahub | Full-featured Community (Apache 2.0) | Undisclosed — contact sales | ”Processing capabilities” based | HA requires Enterprise; Neo4j Enterprise license bundled |
| NetBox | Open-source (Apache 2.0) | NetBox Cloud from ~$500/mo | Node count / hosted instance | Plugin maintenance, self-hosting infra |
| ServiceNow CMDB | None | $50-100+/user/month (ITSM bundle) | Per-user, per-module | Implementation services ($100K+ typical), ongoing admin |
| Nautobot | Open-source (Apache 2.0) | Network to Code consulting | Consulting-based support | Self-hosting, plugin compatibility |
| Port.io | Free tier (limited) | Custom pricing | Per-developer, per-service | Focused on app layer; infra modeling is secondary |
Pricing verified May 7, 2026 from vendor websites and published documentation. Infrahub Enterprise pricing is not publicly available.
Tip: If you are evaluating Infrahub, start with the Community edition and the hosted sandbox at sandbox.infrahub.app. Model your actual infrastructure — not a demo topology — and stress-test the graph queries your AIOps agents would need. This gives you real TCO evidence before entering an Enterprise sales cycle.
Who Should Use a Graph-Based Infrastructure Data Platform
Good fit:
- Organizations running hybrid or multi-cloud infrastructure with complex interdependencies that CMDBs cannot model accurately
- Teams deploying AIOps agents that need relationship-aware context (dependency mapping, blast radius analysis, change impact prediction)
- Network and infrastructure teams managing 1,000+ devices across multiple sites or vendors
- Companies where compliance or regulatory requirements demand an auditable, version-controlled record of infrastructure changes
Not a good fit:
- Small teams with fewer than a few hundred infrastructure assets — a spreadsheet or NetBox instance is likely sufficient
- Organizations whose AIOps needs are limited to log aggregation and alerting (Datadog, New Relic, and Dynatrace handle this without a separate data layer)
- Teams that need an application-layer service catalog more than infrastructure topology mapping — Port.io or Backstage is a better fit
- Companies not ready to invest in data modeling upfront — Infrahub’s schema-first approach requires deliberate design work before you get value
Graph-Based Data Platform Strengths:
- Enables relationship-aware AI reasoning across infrastructure
- Version-controlled data prevents the “CMDB drift” problem
- Scales to complex hybrid environments without flattening topology
- Reduces mean time to resolution when agents can traverse dependency graphs
Graph-Based Data Platform Weaknesses:
- Higher upfront investment in schema design and data migration
- Requires team buy-in to treat infrastructure data as a first-class engineering artifact
- Smaller vendor ecosystem than established CMDB incumbents
- Graph database operations expertise is less common than relational DB skills
Bottom Line
The AIOps industry has a data problem it does not talk about enough. The race to deploy AI agents for incident response, capacity planning, and automated remediation is outpacing the infrastructure data quality those agents require. If your CMDB was last audited during a migration project, your spreadsheets contradict each other, and your scripts parse configs into formats only their author understands — no amount of model tuning will fix the outputs.
OpsMill’s Infrahub is one of the more credible attempts to solve this problem. The graph-native data model, Git-integrated version control, and built-in CI validation address the root causes of bad infrastructure data in ways that bolting AI features onto an existing CMDB does not. The Eurofiber case (five days to fifteen minutes for service deployments) is the kind of concrete result that justifies the migration effort.
The caveats are real: Enterprise pricing is a black box, the ecosystem is early-stage, and the platform is network/infrastructure-focused rather than full-stack. But for AIOps teams hitting a wall because their agents keep acting on stale or incomplete data, fixing the data layer is not optional — it is prerequisite. Whether Infrahub is the right tool depends on your infrastructure complexity and willingness to invest in schema-first data modeling. Start with the free Community edition and find out.
Rating: 3.8 / 5 for enterprise infrastructure teams evaluating AI-ready data platforms. Strong technical foundations and real production traction, held back by pricing opacity and early-stage ecosystem maturity.
Related Articles
Discussion